Game Theory Slides Chapter 5

Download as pdf or txt
Download as pdf or txt
You are on page 1of 84

Chapter 5: Mixed Strategy Nash

Equilibrium

Game Theory:
An Introduction with Step-by-Step Examples

by Ana Espinola-Arredondo and Felix Muñoz-Garcia


Introduction
• We have considered games that had at least one NE
• Prisoner’s Dilemma, Battle of the Sexes, and Chicken games
• But do all games have at least one NE?
• If we restrict players to choose a specific strategy with certainty,
some games may not have a NE.
• This occurs when players do not want to be predictable –playing in the
same way every time they face the game –such as in board games and
professional sports.
Introduction
• Example: Penalty kicks in Soccer
Kicker
Aim left Aim Right
Dive Left 0,0 -10,16
Goalie
Dive Right -10,16 0,0
Matrix 5.1a. Anticoordination Game
• 𝐵𝐵𝐵𝐵𝐺𝐺 𝐿𝐿 = 𝐿𝐿; 𝐵𝐵𝐵𝐵𝐺𝐺 𝑅𝑅 = 𝑅𝑅 and 𝐵𝐵𝐵𝐵𝐾𝐾 𝐿𝐿 = 𝑅𝑅; 𝐵𝐵𝐵𝐵𝐾𝐾 𝑅𝑅 = 𝐿𝐿
• Intuitively, the goalie tries to move in the same direction as the kicker, so
she prevents the latter from scoring. Meanwhile, the kicker seeks to aim to
the opposite location of the goalie to score a goal.
Pure Strategy Nash Equilibrium (psNE)
• Definition. Strategy profiles where players use a specific strategy with 100
percent probability are referred to as “pure-strategy NE”
Kicker
Aim left Aim Right
Dive Left 0,0 -10,16
Goalie
Dive Right -10,16 0,0
Matrix 5.1b. Anticoordination Game underlining best response payoffs

• In the above game, if we restrict players to use a specific strategy, there will be no
mutual best response, and this game will have no NE, i.e., 𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 = ∅ .
• However, if we allow players to randomize, such as playing left with probability =
1⁄ , and right with the remaining probability = 2⁄ , we can find the NE of the
3 3
game. (These strategies are called Mixed Strategy NE (msNE), which we discuss in
later slides.)
Another Example
• Consider a game where police department chooses where to locate most of its police
patrols and, simultaneously, a criminal organization decides where to run its business
• Matrix 5.2 shows that the police seeks to choose the same action as the criminal, while
the latter seeks to miscoordinate by selecting the opposite location as the police patrol to
avoid being caught.
Criminal
Street A Street B
Street A 10,0 -1,6
Police
Street B 0,8 7,-1
Matrix 5.2. Police and Criminal Game
• A similar argument applies to firm monitoring, such as a polluting firm choosing how
many emissions to abate and an environmental protection agency deciding the frequency
of its inspections.
Mixed Strategy
• Consider an individual 𝑖𝑖 with a binary strategy set 𝑆𝑆𝑖𝑖 = 𝐻𝐻, 𝐿𝐿
representing, for instance, a firm choosing between high and low
prices.
• Define a player 𝑖𝑖’s mixed strategy (or randomization) as a probability
distribution over her pure strategies 𝐻𝐻 𝑎𝑎𝑎𝑎𝑎𝑎 𝐿𝐿 , as follows 𝜎𝜎𝑖𝑖 =
𝜎𝜎𝑖𝑖 𝐻𝐻 , 𝜎𝜎𝑖𝑖 𝐿𝐿 , where
𝜎𝜎𝑖𝑖 𝐻𝐻 = 𝑝𝑝 and 𝜎𝜎𝑖𝑖 𝐿𝐿 = 1 − 𝑝𝑝

indicating the probability assigned to each strategy 𝑆𝑆𝑖𝑖 . We require that


𝜎𝜎𝑖𝑖 𝐻𝐻 , 𝜎𝜎𝑖𝑖 𝐿𝐿 ≥ 0 and 𝜎𝜎𝑖𝑖 𝐻𝐻 + 𝜎𝜎𝑖𝑖 𝐿𝐿 = 1.
Mixed Strategy
• Definition. Mixed Strategy. Consider a discrete strategy set 𝑆𝑆𝑖𝑖 =
𝑠𝑠1 , 𝑠𝑠2 , … , 𝑠𝑠𝑚𝑚 where 𝑚𝑚 ≥ 2 denotes number of pure strategies. The
mixed strategy
𝜎𝜎𝑖𝑖 = 𝜎𝜎𝑖𝑖 𝑠𝑠1 , 𝜎𝜎𝑖𝑖 𝑠𝑠2 , . . , 𝜎𝜎𝑖𝑖 𝑠𝑠𝑚𝑚
is a probability distribution over the pure strategies in 𝑆𝑆𝑖𝑖 , with the property
that:
1. 𝜎𝜎𝑖𝑖 𝑠𝑠𝑘𝑘 ≥ 0 for every pure strategy 𝑠𝑠𝑘𝑘 , and
m
2. ∑k=1 σi sk = 1
• When mixed strategy concentrates all probability weight on a pure
strategy 𝜎𝜎𝑖𝑖 𝑠𝑠𝑗𝑗 = 1 while 𝜎𝜎𝑖𝑖 𝑠𝑠𝑘𝑘 = 0 for all 𝑗𝑗 ≠ 𝑘𝑘, it is commonly called a
“degenerated mixed strategy” because, graphically, it collapses to a pure
strategy.
Mixed Strategy
• However, to avoid unnecessary complications, we only use the term “mixed
strategy” to probability distributions over at least two pure strategies.
• As a remark, the above definition can be applied to games where players
choose their strategies from a continuous strategy space, e.g., an output
level so that 𝑠𝑠𝑖𝑖 > 0.
• In this context, player 𝑖𝑖 ′ 𝑠𝑠 probability distribution over her pure strategies in
𝑆𝑆𝑖𝑖 can be represented with a cumulative distribution function
𝐹𝐹𝑖𝑖 : 𝑆𝑆𝑖𝑖 → 0,1
mapping every strategy 𝑠𝑠𝑖𝑖 ∈ 𝑆𝑆𝑖𝑖 into cumulative probability.
• For instance, if 𝑠𝑠𝑖𝑖 denotes firm 𝑖𝑖 ′ 𝑠𝑠 output level, the probability that this
firm produces an output level equal or lower than 𝑠𝑠̅ is 𝐹𝐹𝑖𝑖 𝑠𝑠̅ and, because
𝐹𝐹𝑖𝑖 𝑠𝑠̅ is a probability, it must satisfy 𝐹𝐹𝑖𝑖 𝑠𝑠̅ ∈ 0,1 .
Best Response with Mixed Strategies
• Definition. Best response with Mixed Strategies. Player 𝑖𝑖 ′ 𝑠𝑠 mixed strategy
𝜎𝜎𝑖𝑖 is a best response to her opponents’ mixed strategy 𝜎𝜎−𝑖𝑖 if and only if her
expected utility from 𝜎𝜎𝑖𝑖 satisfies
𝐸𝐸𝐸𝐸𝑖𝑖 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 ≥ 𝐸𝐸𝐸𝐸𝑖𝑖 𝜎𝜎𝑖𝑖′ , 𝜎𝜎−𝑖𝑖 for all 𝜎𝜎𝑖𝑖′ ≠ 𝜎𝜎𝑖𝑖

• Mixed strategy 𝜎𝜎𝑖𝑖 is player 𝑖𝑖 ′ 𝑠𝑠 best response to her opponents’ mixed


strategy 𝜎𝜎−𝑖𝑖 if no other randomization 𝜎𝜎𝑖𝑖′ (potentially including the use of
pure strategies) yields a higher expected utility than 𝜎𝜎𝑖𝑖 does.
• We use expected utility because player 𝑖𝑖 needs to compute her expected
payoff from randomizing over at least two of her pure strategies and,
potentially, her rivals also randomize.
Mixed Strategy Nash Equilibrium (msNE)
• Definition. Mixed Strategy Nash Equilibrium (msNE). A strategy
profile 𝜎𝜎𝑖𝑖∗ , 𝜎𝜎−𝑖𝑖

is a mixed strategy Nash equilibrium if and only if
𝜎𝜎𝑖𝑖∗ = 𝐵𝐵𝐵𝐵𝑖𝑖 𝜎𝜎−𝑖𝑖

for every player 𝑖𝑖.
• Therefore, when player 𝑖𝑖 chooses her equilibrium strategy 𝜎𝜎𝑖𝑖∗ , she is optimally

responding to her opponents’ strategies, 𝜎𝜎−𝑖𝑖 , implying that players are
choosing mutual best responses and, thus, have no incentives to unilaterally
deviate.
• This definition is, then, analogous to that of the pure strategy NE in Chapter 3,
but using mixed strategy 𝜎𝜎𝑖𝑖∗ rather than pure strategy 𝑠𝑠𝑖𝑖∗ .
Goalie Example Revisited
Kicker
Prob. q Prob. 1-q
Aim left Aim Right

Prob. 𝑝𝑝 Dive Left 0,0 -10,16


Goalie Prob. 1-𝑝𝑝 Dive Right -10,16 0,0

Matrix 5.3. Anticoordination Game –including probabilities


For mixed strategy
• The goalie must be indifferent between diving left and right. Her expected utility from diving left
must coincide with that of diving right
𝐸𝐸𝐸𝐸𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = 𝐸𝐸𝐸𝐸𝐺𝐺𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
• Let 𝑝𝑝 and 1 − 𝑝𝑝 denote the probability with which the goalie randomizes, and 𝑞𝑞 and 1 − 𝑞𝑞 be
the probability with which the kicker randomizes
• Intuitively,
• 𝑝𝑝 = 1 means that the goalie dives left with 100 percent probability, whereas 𝑝𝑝 = 0 indicates the opposite
• 0 < 𝑝𝑝 < 1 means the goalie randomizes her diving decision
• Similar argument applies to Kicker’s probability, 𝑞𝑞
Tool 5.1. How to Find msNEs in a two-player
game
1. Focus on player 1. Find her expected utility of choosing one pure
strategy. Repeat for each pure strategy.
2. Set these expected utilities equal to each other.
3. Solve for player 2’s equilibrium mixed strategy, 𝜎𝜎2∗ .
a. If players have two pure strategies, step 2 just entails an equality. Solving for 𝜎𝜎2 in
this context yields a probability, 𝜎𝜎2∗ which should satisfy 0 < 𝜎𝜎2∗ < 1.
b. If players have three pure strategies, step 2 entails several equalities, which gives
rise to a system of two equations and two unknowns. The solution to this system
of equation is, nonetheless, a player 2’s equilibrium mixed strategy, 𝜎𝜎2∗ .
4. Focus now on player 2. Repeat steps 1-3, to obtain player 1’s equilibrium
mixed strategy, 𝜎𝜎1∗ .
Example 5.1. Finding msNE: Goalie
Kicker
Prob. q Prob. 1-q
Aim left Aim Right

Prob. 𝑝𝑝 Dive Left 0,0 -10,16


Goalie Prob. 1-𝑝𝑝 Dive Right -10,16 0,0

Matrix 5.3. Anticoordination Game –including probabilities


• Goalie’s expected utility from diving left:
𝐸𝐸𝐸𝐸𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = 𝑞𝑞 × 0 + 1 − 𝑞𝑞 × −10 = 10𝑞𝑞 − 10
𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
• Goalie’s expected utility from diving right:
𝐸𝐸𝐸𝐸𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = 𝑞𝑞 × −10 + 1 − 𝑞𝑞 × 0 = −10𝑞𝑞
𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
• 𝐸𝐸𝐸𝐸𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = 𝐸𝐸𝐸𝐸𝐺𝐺𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
10𝑞𝑞 − 10 = −10𝑞𝑞 ⇒ 𝑞𝑞 = 1⁄2
Example 5.1. Finding msNE: Kicker
Following the same steps we have,
• Kicker’s expected utility from aiming left:
𝐸𝐸𝐸𝐸𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = 𝑝𝑝 × 0 + 1 − 𝑝𝑝 × 16 = 16 − 16𝑝𝑝
𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
• Kicker’s expected utility from aiming right:
𝐸𝐸𝐸𝐸𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = 𝑝𝑝 × 16 + 1 − 𝑝𝑝 × 0 = 16𝑝𝑝
𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟
• 𝐸𝐸𝐸𝐸𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = 𝐸𝐸𝐸𝐸𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝑒𝑒𝑟𝑟 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
16 − 16𝑝𝑝 = 16𝑝𝑝 ⇒ 𝑝𝑝 = 1⁄2

1 1 1 1
• 𝑚𝑚𝑚𝑚𝑚𝑚𝐸𝐸 = 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿, 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 ; 𝐴𝐴𝐴𝐴𝐴𝐴 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿, 𝐴𝐴𝐴𝐴𝐴𝐴 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
2 2 2 2

𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾
• Remember that players do not need to randomize with the same probability. They only did in this case because payoffs
are symmetric in Matrix 5.3.
Graphical Visualization: msNE
• How to graphically represent the best response of each player?
• Let’s start with the Goalie
• Goalie chooses to dive left if:
𝐸𝐸𝐸𝐸𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 > 𝐸𝐸𝐸𝐸𝐺𝐺𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
10𝑞𝑞 − 10 > −10𝑞𝑞 ⇒ 𝑞𝑞 > 1�2
• Mathematically, this means that, for all 𝑞𝑞 > 1⁄2, the goalie chooses to dive
left (i.e., 𝑝𝑝 = 1).
• In contrast, for all 𝑞𝑞 < 1⁄2, the goalie responds by diving right (i.e., 𝑝𝑝 = 0).
Graphical Visualization: msNE

We can summarize the BRF of Goalie as:


1
𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝑖𝑖𝑖𝑖 𝑞𝑞 >
2
1
𝐵𝐵𝐵𝐵𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝑞𝑞 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿, 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑖𝑖𝑖𝑖𝑖𝑖 = , 𝑎𝑎𝑎𝑎𝑎𝑎
2
1
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑖𝑖𝑖𝑖𝑖𝑖 <
2

Figure 5.1 depicts this best response


function
Graphical Visualization: msNE
• For Kicker
• Kicker aims left if:
𝐸𝐸𝐸𝐸𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 > 𝐸𝐸𝐸𝐸𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝑒𝑒𝑟𝑟 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
16 − 16𝑝𝑝 > 16 ⇒ 𝑝𝑝 < 1�2
• Mathematically, this means that, for all p < 1⁄2, the kicker chooses to aim
left (i.e., q = 1).
• In contrast, for all p > 1⁄2, the kicker aims right (i.e., q = 0).
Graphical Visualization: msNE

We can summarize the BRF of Kicker as:


1
𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝑖𝑖𝑖𝑖 𝑝𝑝 <
2
1
𝐵𝐵𝐵𝐵𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝑞𝑞 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿, 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑖𝑖𝑖𝑖𝑖𝑖 = , 𝑎𝑎𝑎𝑎𝑎𝑎
2
1
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑖𝑖𝑖𝑖𝑖𝑖 >
2

Figure 5.2 depicts this best response


function
Graphical Visualization: msNE
• Putting together their
response
• Figure 5.3 superimposes the
goalie’s and the kicker’s best
response functions
• The player’s best responses
only cross each other at one
point, where 𝑝𝑝 = 𝑞𝑞 = 1⁄2
Some Lessons
• Indifference.
• If it is optimal to randomize over a collection of pure strategies, then a player
receives the same expected payoff from each of those pure strategies.
• When analyzing player 𝑖𝑖, ignore her probability to randomize
• We make player 𝑖𝑖 indifferent between two (or more) of her pure strategies, we write
player 𝑖𝑖’s expected utility from choosing pure strategy, 𝑠𝑠𝑖𝑖 , 𝐸𝐸𝐸𝐸𝑖𝑖 𝑠𝑠𝑖𝑖 as a function of
her rival’s randomization
• Never use strictly dominated strategies
• If a pure strategy is strictly dominated, then such a pure strategy cannot be part of
psNE or a msNE.
• Odd number of equilibria
• In almost all finite games (games with a finite set of players and available actions),
there is a finite and odd number of equilibria
Extensions: Mixed Strategy Equilibria in
Games with 𝑘𝑘 ≥ 3 pure strategies
• Example: Rock-Paper-Scissors game
Player 2 Player 2
R P S R P S
R 0,0 -1,1 1,-1 R 0,0 -1,1 1,-1
Player 1 P 1,-1 0,0 -1,1 Player 1 P 1,-1 0,0 -1,1
S -1,1 1,-1 0,0 S -1,1 1,-1 0,0
Matrix 5.4. Rock-Paper-Scissors Game Matrix 5.4a. Rock-Paper-Scissors Game –
Underlining best response payoffs
• No psNE exists in this game
• Player’s payoffs are symmetric entailing that they randomize with the same
probabilities: 𝑟𝑟, 𝑝𝑝, and 1 − 𝑟𝑟 − 𝑝𝑝, where 𝑟𝑟 denotes the probability that every
player 𝑖𝑖 chooses Rock, 𝑝𝑝 represents the probability she selects Paper, and
1 − 𝑟𝑟 − 𝑝𝑝 is the probability she plays Scissors
Example 5.2. Finding msNE when players have
three pure strategies
• If player 𝑖𝑖 randomizes, she must be indifferent between all her three pure strategies
𝐸𝐸𝐸𝐸𝑖𝑖 𝑅𝑅 = 𝐸𝐸𝐸𝐸𝑖𝑖 𝑃𝑃 = 𝐸𝐸𝐸𝐸𝑖𝑖 𝑆𝑆
• We separately find player 1’s expected utility from each pure strategy
• To compute 𝐸𝐸𝐸𝐸1 𝑅𝑅 , focus at the top row of Matrix 5.4a where:
• If player 2 chooses Rock, which happens with probability 𝑟𝑟, player 1 earns a payoff of zero,
• If player 2 chooses Paper, which happens with probability 𝑝𝑝, player 1 earns a payoff of -1, and
• If player 2 chooses Scissors, which happens with probability 1 − 𝑟𝑟 − 𝑝𝑝, player 1 earns a payoff of 1
• In summary, player 1’s expected utility from Rock is:

𝐸𝐸𝐸𝐸1 𝑅𝑅 = 𝑟𝑟0 + 𝑝𝑝 −1 + 1 − 𝑟𝑟 − 𝑝𝑝 1 = 1 − 𝑟𝑟 − 2𝑝𝑝

• Working in a similar way:


𝐸𝐸𝐸𝐸1 𝑃𝑃 = 𝑟𝑟1 + 𝑝𝑝𝑝 + 1 − 𝑟𝑟 − 𝑝𝑝 −1 = 2𝑟𝑟 + 𝑝𝑝 − 1

𝐸𝐸𝐸𝐸1 𝑆𝑆 = 𝑟𝑟 −1 + 𝑝𝑝𝑝 + 1 − 𝑟𝑟 − 𝑝𝑝 0 = 𝑝𝑝 − 𝑟𝑟
Example 5.2. Finding msNE when players have
three pure strategies
• We can now set the expected utilities from Rock, Paper, and Scissors equal to each other.
• First, from 𝐸𝐸𝐸𝐸1 𝑅𝑅 = 𝐸𝐸𝐸𝐸1 𝑆𝑆 we find that:
1
1 − 𝑟𝑟 − 2𝑝𝑝 = 𝑝𝑝 − 𝑟𝑟 ⇒ 𝑝𝑝 =
3
• Second, 𝐸𝐸𝐸𝐸1 𝑃𝑃 = 𝐸𝐸𝐸𝐸1 𝑆𝑆 we find that:
1
2𝑟𝑟 + 𝑝𝑝 − 1 = 𝑝𝑝 − 𝑟𝑟 ⇒ 𝑟𝑟 =
3
1
• Therefore 1 − r − 𝑝𝑝 = .
3
• We can then summarize the msNE of this game as that, every player 𝑖𝑖, randomizes
according to the mixed strategy
1 1 1
𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝑟𝑟 ∗ , 𝑝𝑝∗ , 1 − 𝑝𝑝∗ − 𝑟𝑟 ∗ = , ,
3 3 3
which means that every player assigns the same probability weight to each of her pure strategies.
Finding Mixed Strategy Equilibria in games
with 𝑁𝑁 ≥ 2 players
“Snob Effect” game
• Consider that every player 𝑖𝑖 independently and simultaneously chooses
between alternatives 𝑋𝑋 and 𝑌𝑌, where 𝑋𝑋 can be interpreted as the snob
option, while 𝑌𝑌 is the conforming option.
• If she is the only player choosing 𝑋𝑋, she earns a payoff of 𝑎𝑎 (she is the ”cool
girl” in the group), but a payoff of 𝑏𝑏 otherwise (if anyone else chooses X too)
• When she chooses 𝑌𝑌, she earns a payoff of 𝑐𝑐 regardless of how many other
players choose 𝑋𝑋.
• Payoffs satisfy 𝑎𝑎 > 𝑐𝑐 > 𝑏𝑏.
Finding Mixed Strategy Equilibria in games
with 𝑁𝑁 ≥ 2 players
“Snob Effect” game
• Before extending the game to N players, let us consider the two-player version.
Player 2
Prob. p Prob. 1-p
X Y
Prob. 𝑝𝑝 X b, b a, c
Player
1 Prob. 1-𝑝𝑝 Y c, a c, c

• Two pure strategy NEs: (X,Y) and (Y,X), as in anticoordination games.


• What about msNE?
• Player 1 is indifferent between X and Y if and only if 𝑝𝑝𝑝𝑝 + 1 − 𝑝𝑝 𝑎𝑎 = 𝑝𝑝𝑝𝑝 + 1 − 𝑝𝑝 𝑐𝑐, which
𝑎𝑎−𝑐𝑐
yields 𝑝𝑝∗ = . (Same probability for player 2, since payoffs are symmetric.)
𝑎𝑎−𝑏𝑏
• Probability 𝑝𝑝∗ is positive and smaller than one since 𝑎𝑎 > 𝑐𝑐 > 𝑏𝑏. Check!
• 𝑝𝑝∗ is increasing in 𝑎𝑎 − 𝑐𝑐, but decreasing in 𝑎𝑎 − 𝑏𝑏. Intuition.
Finding Mixed Strategy Equilibria in games
with 𝑁𝑁 ≥ 2 players
“Snob Effect” game
• We can now extend the game to N players.
• In this setting, we seek to identify a symmetric msNE where every player 𝑖𝑖
chooses 𝑋𝑋 with probability 𝑝𝑝.
• Therefore, we need to:
1. Find the expected utility from 𝑋𝑋 and from 𝑌𝑌, and
2. Set these expected utilities equal to each other to obtain the equilibrium probability
𝑝𝑝∗
Game continued
• Expected Utility from 𝑿𝑿. When player 𝑖𝑖 chooses 𝑋𝑋, his expected utility is

𝐸𝐸𝐸𝐸𝑖𝑖 𝑋𝑋 = (1 − 𝑝𝑝)𝑁𝑁−1 𝑎𝑎 + 1 − 1 − 𝑝𝑝 𝑁𝑁−1 𝑏𝑏


𝑁𝑁𝑁𝑁 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 𝐴𝐴𝐴𝐴 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒
𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑋𝑋 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑋𝑋
To understand the probabilities of each event, recall that the probability with
which every player chooses 𝑋𝑋 is 𝑝𝑝, so the probability with which she plays 𝑌𝑌
is 1 − 𝑝𝑝, implying that all other 𝑁𝑁 − 1 players (everyone but player 𝑖𝑖) choose
alternative 𝑌𝑌 with probability
1 − 𝑝𝑝 × 1 − 𝑝𝑝 × ⋯ × (1 − 𝑝𝑝) = (1 − 𝑝𝑝)𝑁𝑁−1
𝑁𝑁−1 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡
Game continued
• Expected Utility from 𝒀𝒀. When player 𝑖𝑖 chooses 𝑌𝑌, she earns a payoff
of 𝑐𝑐 with certainty, entailing that 𝐸𝐸𝐸𝐸𝑖𝑖 𝑌𝑌 = 𝑐𝑐.
• Indifference condition. If player 𝑖𝑖 randomizes between 𝑋𝑋 and 𝑌𝑌, she
must be indifferent, thus earning the same expected payoff from each
pure strategy, 𝐸𝐸𝐸𝐸𝑖𝑖 𝑋𝑋 = 𝐸𝐸𝐸𝐸𝑖𝑖 𝑌𝑌 , which means:
(1 − 𝑝𝑝)𝑁𝑁−1 𝑎𝑎 + 1 − 1 − 𝑝𝑝 𝑁𝑁−1 𝑏𝑏 = 𝑐𝑐
Rearranging, yields
1
𝑐𝑐 − 𝑏𝑏 𝑁𝑁−1
𝑝𝑝∗ =1−
𝑎𝑎 − 𝑏𝑏
Comparative Statics
• For instance, when 𝑎𝑎 = 2, 𝑏𝑏 = 0, 𝑎𝑎𝑎𝑎𝑎𝑎 𝑐𝑐 =
1, the probability simplifies to
1
𝑝𝑝∗ = 1 − 1�2 𝑁𝑁−1

which decreases in the number of players, 𝑁𝑁.


• Intuitively, as the population grows, the
probability that someone selects
alternative 𝑋𝑋 increases, driving each player
to individually decrease their probability of
choosing the snob option 𝑋𝑋, opting instead
for conforming option
1
𝑌𝑌, with probability
1 − 𝑝𝑝∗ = 1⁄2 𝑁𝑁−1 , which increase in 𝑁𝑁
1

Comparative Statics ∗
𝑝𝑝 = 1 −
𝑐𝑐 − 𝑏𝑏
𝑎𝑎 − 𝑏𝑏
𝑁𝑁−1

• We can also examine how the above expression 𝑝𝑝∗ changes when:
• Players are more attracted to the snob option, 𝑎𝑎 − 𝑏𝑏, which captures the
payoff gain of snob option relative to the conforming option.
• An increase in 𝑎𝑎 − 𝑏𝑏 increases the probability of choosing 𝑋𝑋.
• In contrast, an increase in 𝑐𝑐 − 𝑏𝑏, which measures the payoff loss of
not being the only individual choosing 𝑋𝑋, decreases the probability of
choosing 𝑋𝑋.
Strictly Competitive Games
• Definition. Strictly Competitive Games. A two-player game is strictly
competitive if, for every two strategy profiles 𝑠𝑠 and 𝑠𝑠 ′ ,
• if 𝑢𝑢1 𝑠𝑠 > 𝑢𝑢1 𝑠𝑠 ′ then 𝑢𝑢2 𝑠𝑠 < 𝑢𝑢2 𝑠𝑠 ′ ; and
• if 𝑢𝑢1 𝑠𝑠 = 𝑢𝑢1 𝑠𝑠 ′ , then 𝑢𝑢2 𝑠𝑠 = 𝑢𝑢2 𝑠𝑠 ′ .

• Intuitively:
• If player 1 prefers strategy profile 𝑠𝑠 to 𝑠𝑠 ′ , then player 2 has the opposite preference
order: preferring 𝑠𝑠 ′ over 𝑠𝑠; and
• if player 1 is indifferent between 𝑠𝑠 and 𝑠𝑠 ′ , player 2 must also be indifferent between
these two strategy profiles.
• Example: The penalty kicks game is an example of a strictly competitive
game where we can test the above definition (next slide).
Strictly Competitive Games
Kicker
Aim left (l) Aim Right (r)
Dive Left (L) 0,0 -10,16
Goalie
Dive Right (R) -10,16 0,0

• Comparing 𝐿𝐿, 𝑙𝑙 and 𝐿𝐿, 𝑟𝑟 , we see that the goalie prefers the former, since 0 > −10,
while the kicker prefers the latter because 0 < 16.
• Comparing 𝐿𝐿, 𝑙𝑙 and 𝑅𝑅, 𝑙𝑙 , we find that the goalie prefers the former, since 0 > −10,
while the kicker prefers the latter because 0 < 16.
• Comparing 𝐿𝐿, 𝑙𝑙 and 𝑅𝑅, 𝑟𝑟 , we see that the goalie is indifferent, and so is the kicker,
both players earning a payoff of zero in both strategy profiles.
• Comparing 𝑅𝑅, 𝑙𝑙 and 𝐿𝐿, 𝑟𝑟 , we find that the goalie is indifferent between these two
strategy profiles, earning −10 in both of them. A similar argument applies to the kicker,
who earns a payoff of 16 in both strategy profiles.
• We can confirm the definition of strictly competitive games (i.e.,′.
opposite preferences of
players 1 and 2) holds for every two strategy profiles, 𝑠𝑠 and 𝑠𝑠
Games that are not strictly competitive
• A two-player game is not strictly competitive if, for at least two strategy
profiles, 𝑠𝑠 and 𝑠𝑠 ′ , every player 𝑖𝑖 ′ 𝑠𝑠 utility satisfies 𝑢𝑢𝑖𝑖 𝑠𝑠 > 𝑢𝑢𝑖𝑖 𝑠𝑠 ′ .
Criminal

• Example Street A Street B


Street A 10,0 -1,6
Police
Street B 0,8 7,-1
Matrix 5.6. Police and Criminal Game

• Comparing strategy profiles 𝐴𝐴, 𝐴𝐴 and 𝐵𝐵, 𝐵𝐵 , along the main diagonal, we can see
that the police prefers 𝐴𝐴, 𝐴𝐴 to 𝐵𝐵, 𝐵𝐵 , since her payoff satisfies 10 > 7.
• Similar argument applies for the criminal, as her payoff satisfies 0 > −1.
• Because we found that players’ preferences over strategy profiles are aligned, rather
than misaligned, we can already claim that the game is not strictly competitive
without having to compare other pairs of strategy profiles.
Constant-sum Games
• Definition. Constant-sum games. A two-player game is a constant-
sum game if, for every strategy profile 𝑠𝑠, player’s payoffs satisfy
𝑢𝑢1 𝑠𝑠 + 𝑢𝑢2 𝑠𝑠 = 𝐾𝐾, where 𝐾𝐾 > 0 is a constant.

• Then, players’ payoffs must add up to the same constant across all
cells in the matrix.
• If, instead, players’ payoffs add up to a different number in at least
one of the cells, then we can claim that the game is not constant sum.
• It can still be strictly competitive, but not constant sum.
Constant-sum Games
Counterexample:
Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 10,0 9,3
Player 1
𝐷𝐷 9,3 10,0
Matrix 5.7. A strictly competitive game that is non constant-sum

• The game is strictly competitive (check as practice).


• It is not a constant-sum game since players payoff in strategy profiles
like 𝑈𝑈, 𝑙𝑙 and 𝐷𝐷, 𝑟𝑟 add up to 10, while those strategy profiles 𝑈𝑈, 𝑟𝑟
and 𝐷𝐷, 𝑙𝑙 add up to 12.
Constant-sum Games

• Constant-sum games are always strictly competitive:


• Condition 𝑢𝑢1 𝑠𝑠 + 𝑢𝑢2 𝑠𝑠 = 𝐾𝐾 can be rewritten as 𝑢𝑢1 𝑠𝑠 = 𝐾𝐾 − 𝑢𝑢2 𝑠𝑠 .
• Then, if player 1’s payoff increases when moving from 𝑠𝑠 to 𝑠𝑠′, then player 2’s
payoff must decrease.

• We now introduce a special class of constant-sum games, those in


which K=0, called zero-sum games.
Zero-sum Games
• Definition. Zero-sum games. A two-player game is a zero-sum game if, for every
strategy profile 𝑠𝑠, player’s payoffs satisfy
𝑢𝑢1 𝑠𝑠 + 𝑢𝑢2 𝑠𝑠 = 0.
• Alternatively, condition 𝑢𝑢1 𝑠𝑠 + 𝑢𝑢2 𝑠𝑠 = 0 can be expressed as 𝑢𝑢1 𝑠𝑠 = −𝑢𝑢2 𝑠𝑠 .

• Intuitively, every dollar that player 1 earns comes from the same dollar that player 2 loses
and vice versa.
Player 2
𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻 1,-1 -1,1
Player 1
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 -1,1 1,-1
Matrix 5.8. Matching Pennies Game
• Matching pennies game is zero-sum game. Rock-paper-scissors is another example.
• Specifically, in Matrix 5.8, we have that either 1 + −1 = 0 or −1 + 1 = 0.
Security Strategies
• Definition. Security Strategies. In a two-player game, player 𝑖𝑖 ′ 𝑠𝑠
security strategy, 𝑖𝑖, solves
max m𝑖𝑖𝑖𝑖 𝑢𝑢𝑖𝑖 𝜎𝜎𝑖𝑖 , 𝜎𝜎𝑗𝑗
𝜎𝜎𝑖𝑖 𝜎𝜎𝑗𝑗
• Consider the “worst-case scenario” 𝑤𝑤𝑖𝑖 𝜎𝜎𝑖𝑖 = min 𝑢𝑢𝑖𝑖 𝜎𝜎𝑖𝑖 , 𝜎𝜎𝑗𝑗
𝜎𝜎𝑗𝑗
• Player 𝑖𝑖 anticipates that player 𝑗𝑗 chooses her strategy 𝜎𝜎𝑗𝑗 to maximize her own
payoff, which entails minimizing 𝑖𝑖’s payoff, 𝑢𝑢𝑖𝑖 𝜎𝜎𝑖𝑖 , 𝜎𝜎𝑗𝑗 .
• This is because players interact in a strictly competitive game.
• Player 𝑖𝑖 then chooses her strategy 𝜎𝜎𝑖𝑖 to maximize the payoff across all worst-
case scenarios.
• Intuitively, player 𝑖𝑖 seeks to find the strategy 𝜎𝜎𝑖𝑖 that provides her with the
”best of the worst” payoffs, as represented with the max-min problem.
• This explains why security strategies are sometimes known as max-min strategies.
Tool 5.2. How to find security strategies in a
two-player game
1. Find the expected utility of player 1’s randomization, fixing player 2’s
strategy.
2. Repeat step 1 until you considered all strategies of player 2, fixing one at
a time.
3. “Min” part. Find the lower envelope of player 1’s expected utility. That is,
for each strategy 𝜎𝜎1 , find the lowest expected utility that player 1 earns.
4. “Max” part. Find the highest expected utility of the lower envelope
identified in step 3, and the corresponding strategy 𝜎𝜎1 . This is player 1’s
security strategy, 𝜎𝜎1𝑠𝑠𝑠𝑠𝑠𝑠 .
5. To find the security strategy for player 2, follow a similar process in steps
1-4 above.
Example 5.3. Finding Security Strategies
Example
Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 10,0 9,3
Player 1
𝐷𝐷 9,3 10,0
Matrix 5.9. A Strictly Competitive Game that is non constant-sum

To find the security strategy for player 1, we follow the next steps:
1. We find player 1’s expected utility of randomizing between 𝑈𝑈 and
𝐷𝐷, with associated probabilities 𝑝𝑝 and 1 − 𝑝𝑝, respectively. First, we
fix player 2’s strategy at column 𝑙𝑙, which yields:
𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑙𝑙 = 𝑝𝑝 × 10 + 1 − 𝑝𝑝 × 9 = 9 + 𝑝𝑝
Example 5.3. Finding Security Strategies
2. We now find her expected utility of
randomizing, but fixing player 2’s strategy
at column 𝑟𝑟, as follows:
𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑟𝑟 = 𝑝𝑝 × 9 + 1 − 𝑝𝑝 × 10 = 10 − 𝑝𝑝

3. To find the lower envelope of the previous


two expected utilities, we can depict each
line as a function of 𝑝𝑝, as we do in Figure
5.5. The lower envelope is the segment 9 +
1
𝑝𝑝 for all 𝑝𝑝 ≤ , but segment 10 − 𝑝𝑝
2
otherwise.
Example 5.3. Finding Security Strategies
4. Among all points in the lower envelope, player 1 enjoys the highest
1
utility at 𝑝𝑝 = , which yields an expected payoff 𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑙𝑙 = 9 +
2
1
= 9.5, as illustrated in Figure 5.5 by the height of the crossing
2
point between 𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑙𝑙 and 𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑟𝑟 . This is player 1’s security
𝑆𝑆𝑆𝑆𝑆𝑆 1
strategy, 𝑝𝑝 = .
2
5. Following the same steps for player 2, we find that, since payoffs
𝑆𝑆𝑆𝑆𝑆𝑆 1
are symmetric, her security strategy is 𝑞𝑞 = .
2
Security Strategies and NE
• At this point, you may be wondering about the relationship between
security strategies and msNE.
• We obtain the same equilibrium result from both solution concepts,
but only for two-player strictly competitive games.
• Consider the previous example from Matrix 5.9:
𝑠𝑠𝑠𝑠𝑠𝑠 𝑠𝑠𝑠𝑠𝑠𝑠
1 1
𝑝𝑝 , 𝑞𝑞 = ,
2 2
Let us now confirm that the msNE produces the same result.
Example 5.4. Solved by msNE
• Assuming that player 1 randomizes between 𝑈𝑈 and 𝐷𝐷 with
probabilities 𝑝𝑝 and 1 − 𝑝𝑝, respectively, and player 2 mixes 𝑙𝑙 and 𝑟𝑟
with probabilities 𝑞𝑞 and 1 − 𝑞𝑞, respectively,
• We find that player 1’s expected utility from choosing 𝑈𝑈 is:
𝐸𝐸𝐸𝐸1 𝑈𝑈 = 𝑞𝑞 × 10 + 1 − 𝑞𝑞 × 9 = 9 + 𝑞𝑞
• Similarly, player 1’s expected utility from choosing 𝐷𝐷 is:
𝐸𝐸𝐸𝐸1 𝐷𝐷 = 𝑞𝑞 × 9 + 1 − 𝑞𝑞 × 10 = 10 − 𝑞𝑞
• Therefore, player 1 randomizes between 𝑈𝑈 and 𝐷𝐷 when she is indifferent
between these two pure strategies 𝐸𝐸𝐸𝐸1 𝑈𝑈 = 𝐸𝐸𝐸𝐸1 𝐷𝐷 , which entails
1
9 + 𝑞𝑞 = 10 − 𝑞𝑞 ⇒ 𝑞𝑞 =
2
Example 5.4. Solved by msNE
• Player 2’s expected utilities
• 𝐸𝐸𝐸𝐸2 𝑙𝑙 = 𝑝𝑝 × 0 + 1 − 𝑝𝑝 × 3 = 3 − 3𝑝𝑝, and
• 𝐸𝐸𝐸𝐸2 𝑟𝑟 = 𝑝𝑝 × 3 + 1 − 𝑝𝑝 × 0 = 3𝑝𝑝
1
• 𝐸𝐸𝐸𝐸2 𝑙𝑙 = 𝐸𝐸𝐸𝐸2 𝑟𝑟 ⇒ 3 − 3𝑝𝑝 = 3𝑝𝑝 ⇒ 𝑝𝑝 =
2

• Summarizing, we can claim that the msNE of this game is 𝑝𝑝, 𝑞𝑞 =


1 1
, , which coincides with the security strategies we found in
2 2
example 5.3.
Example 5.5. Security strategies and msNE
yield different equilibrium outcomes
Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 3,5 -1,1
Player 1
𝐷𝐷 2,6 1,2
Matrix 5.10. A Game that is not strictly competitive

• The above game is not strictly competitive. We can find strategy


profile where players’ interests are aligned; both players prefer, for
instance, 𝑈𝑈, 𝑙𝑙 to 𝐷𝐷, 𝑟𝑟 .
• Since the game is not strictly competitive, we can expect that security
strategies may produce a different equilibrium prediction than msNE.
Example 5.5. Security strategies and msNE
yield different equilibrium outcomes
• For player 1:
• When player 2 chooses 𝑙𝑙, player 1’s expected payoff from randomizing
between 𝑈𝑈 and 𝐷𝐷 with probabilities 𝑝𝑝 and 1 − 𝑝𝑝 respectively,
𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑙𝑙 = 𝑝𝑝 × 3 + 1 − 𝑝𝑝 × 2 = 2 + 𝑝𝑝
• When player 2 chooses 𝑟𝑟, player 1’s expected utility is
𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑟𝑟 = 𝑝𝑝 × (−1) + 1 − 𝑝𝑝 × 1 = 1 − 2𝑝𝑝
Example 5.5. Security strategies and msNE
yield different equilibrium outcomes
• 𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑙𝑙 lies above 𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑟𝑟 for all
𝑝𝑝 ∈ 0,1 .
• This means that the lower envelope
coincides with 𝐸𝐸𝐸𝐸1 𝑝𝑝|𝑟𝑟 = 1 − 2𝑝𝑝 for
all values of 𝑝𝑝.
• The highest point of this lower envelope
occurs at 𝑝𝑝 = 0, so player 1 assigns no
probability weight to 𝑈𝑈 or, alternatively,
that she plays 𝐷𝐷 in pure strategies.
• This means that 𝐷𝐷 is player 1’s security
strategy.
Security strategies and msNE yield different
equilibrium outcomes
Similarly, for player 2
• 𝐸𝐸𝐸𝐸2 𝑞𝑞|𝑈𝑈 = 𝑞𝑞 × 5 + 1 − 𝑞𝑞 × 1 = 1 + 4𝑞𝑞, and
• 𝐸𝐸𝐸𝐸2 𝑞𝑞|𝐷𝐷 = 𝑞𝑞 × 6 + 1 − 𝑞𝑞 × 2 = 2 + 4𝑞𝑞.

• Since 𝐸𝐸𝐸𝐸2 𝑞𝑞|𝑈𝑈 < 𝐸𝐸𝐸𝐸2 𝑞𝑞|𝐷𝐷 for all values of 𝑞𝑞,we can claim that 𝑈𝑈 is the
lower envelope.
• We can, then, notice that the highest point of 1 + 4𝑞𝑞 occurs at 𝑞𝑞 = 1,
meaning that player 1 puts full probability weight on 𝑙𝑙, which becomes his
security strategy.
In summary, the security strategy profile in this game is 𝐷𝐷, 𝑙𝑙 .
Example contd. & solving by msNE
• For msNE
• We can facilitate our analysis by noticing that strategy 𝑙𝑙 strictly dominates 𝑟𝑟 since it
yields a strictly higher payoff than 𝑟𝑟 regardless of the row that player 1 chooses
5 > 1 and 6 > 2 .
• We know players put no probability weight in strictly dominated strategies, so we can
delete column 𝑟𝑟 from the matrix and obtain:
Player 2
𝑙𝑙
𝑈𝑈 3,5
Player 1
𝐷𝐷 2,6
Matrix 5.10. A Game that is not strictly competitive – After deleting column r
• Turning now to player 1, we do not need to consider his randomization since, at this
point, he has a clear best response to 𝑙𝑙, 𝑈𝑈. Therefore, the psNE (no msNE) is 𝑈𝑈, 𝑙𝑙 .
• This equilibrium outcome does not coincide with the security strategy profile 𝐷𝐷, 𝑙𝑙 .
Correlated Equilibrium
Player 2
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
• Example
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 6,6 2,7
Player 1
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 7,2 0,0
Matrix 5.11. Modified Chicken game

• By underlining best response payoffs, we can find two psNEs: 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 and
𝑆𝑆𝑡𝑡𝑡𝑡𝑡𝑡, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 .
2
• The game also has msNE, where player 1 chooses 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 with probability 𝑝𝑝 = , and the same
2 3
probability applies to player 2, q= (since payoffs are symmetric).
3
• In this msNE, Player 1’s expected utility in equilibrium is
2 2 1 1 2 1 28 14 14
𝐸𝐸𝐸𝐸1 𝜎𝜎 ∗ = 6+ 2 + 7+ 0 = + = ≅ 4.67
3 3 3 3 3 3 9 9 3

𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 1 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 1 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆


And a similar expected payoff accrues to player 2.
Correlated Equilibrium
• Natural question: Can players reach a higher expected payoff if, instead,
they rely on a probability distribution, such as a coin toss, that each player
privately observes before playing the game, and that informs the player
about which action to choose?
• Intuitively, the probability distribution can be interpreted as an external
“recommender” who:
• First, draws one strategy profile, 𝑠𝑠 = 𝑠𝑠𝑖𝑖 , 𝑠𝑠−𝑖𝑖 , such as one cell in Matrix 5.11.
• Second, the recommender makes recommendation 𝑠𝑠𝑖𝑖 to player 𝑖𝑖, without informing
her of the recommendation 𝑠𝑠−𝑖𝑖 that her rivals receive.
• Definition. Correlated Equilibrium. A probability distribution over strategy
profiles is a correlated equilibrium if every player 𝑖𝑖 follows his
recommendation, 𝑠𝑠𝑖𝑖 .
• Intuitively, a probability distribution over strategy profiles is stable in the sense that
every player 𝑖𝑖 has no incentives to unilaterally deviate from the recommendation, 𝑠𝑠𝑖𝑖 .
• For simplicity, we first examine public signals, then privately observed
signals.
Example 5.6. Correlated Equilibrium with Public Signals
• Consider the game in Matrix 5.11 and assume that players observe a public signal
that assigns probability:
• 𝛼𝛼 to one of the psNEs in this game, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , and
• 1 − 𝛼𝛼 to the other psnE, 𝑆𝑆𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 .

• A public signal could be:


• A traffic light,
• Coin toss,
• Dice toss,
or any other stochastic mechanism that players agree on before starting the game,
that yields this probability distribution (summarized in Matrix 5.12).
Player 2 Player 2
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 6,6 2,7 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 0 1 − 𝛼𝛼
Player 1 Player 1
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 7,2 0,0 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝛼𝛼 0
Matrix 5.12. Correlated equilibrium with public signals –
Matrix 5.11. Modified Chicken game Probability of each strategy profile
Example 5.6. Correlated Equilibrium with Public Signals
Player 2 Player 2
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 6,6 2,7 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 0 1 − 𝛼𝛼
Player 1 Player 1
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 7,2 0,0 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝛼𝛼 0
Matrix 5.11. Modified Chicken game Matrix 5.12. Correlated equilibrium with public signals –
Probability of each strategy profile
• In this context, player 1 does not have incentives to deviate.
• Upon observing 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , payoff from following the recommendation is 7, and
that of unilaterally deviating to 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆e decreases to 6.
• Similarly, upon observing 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , payoff is 2, but decreases to 0 if she
unilaterally deviates.
• By symmetry, the same argument applies to player 2.
• As a consequence, a continuum of correlated equilibria can be sustained, where players
alternate between the two psNEs of the game with probabilities 𝛼𝛼 and 1 − 𝛼𝛼,
respectively.
Example 5.7. Correlated Equilibrium with Private Signals
• Consider the game in Matrix 5.11 again, where the recommendations assign the same probability
weight to 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , 𝑆𝑆𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , and 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , as summarized in Matrix 5.13:
Player 2
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 1� 1�
3 3
Player 1
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 1� 0
3
Matrix 5.13. Correlated equilibrium with private signals – Probability of each strategy profile
• Bottom row. Intuitively, if player 1 receives the recommendation of 𝑆𝑆𝑡𝑡𝑡𝑡y, she knows that the only
strategy profile recommended by the public signal is 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 .
• Top row. However, if she receives the recommendation of 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑒𝑒, she knows that player 2 may
have received:
1
• The same recommendation, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆e, with probability 3
1 1 =1⁄2 , or
+
3 13

• The opposite recommendation, 𝑆𝑆𝑡𝑡𝑡𝑡𝑡𝑡, with probability 3


1 1 =1⁄2.
+
3 3

• A similar interpretation applies to the recommendations that player 2 receives.


Example 5.7. Correlated Equilibrium with Private Signals
We can now show that player 1 does not have incentives to deviate from this
recommendation profile:
• If he receives the recommendation of 𝑆𝑆𝑡𝑡𝑡𝑡𝑡𝑡, his payoff is 7 at 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆, 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , which
she cannot improve by unilaterally deviating to 𝑆𝑆𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 (earning only 6).
• If instead, she receives the recommendation of 𝑆𝑆𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤, her expected payoff is
1 1
6 + 2 = 4,
2 2
which he cannot improve by unilaterally deviating to 𝑆𝑆𝑡𝑡𝑡𝑡𝑡𝑡, as that only yields
1 1
7 + 0 = 3.5.
2 2
• Since payoffs are symmetric, a similar argument applies to player 2, making the above
recommendation stable.
• We can then say that the recommendation profile can be sustained as a correlated
equilibrium, with expected payoff:
1 1 1
6+ 2+ 7=5
3 3 3
which exceeds that in the msNE of the game, 4.67.
Existence of Correlated equilibrium
• Finally, note that every psNE, 𝑠𝑠 ∗ = 𝑠𝑠𝑖𝑖∗ , 𝑠𝑠−𝑖𝑖

can be defined as a (trivial) correlated
equilibrium:
• where the probability distribution recommends player 𝑖𝑖 to choose 𝑠𝑠𝑖𝑖∗ with probability 1.
• That is, every player is recommended to play as she would under the psNE.
• A similar argument applies to msNE 𝜎𝜎 ∗ = 𝜎𝜎𝑖𝑖∗ , 𝜎𝜎−𝑖𝑖

, where the probability
distribution recommends:
• Player 𝑖𝑖 to randomize according to the same mixed strategy that she uses in the msNE of the
game.
• In other words, if 𝜎𝜎 ∗ is a NE, it must also be a correlated equilibrium, but the
converse is not necessarily true.
𝜎𝜎 ∗ is a NE ⇒ 𝜎𝜎 ∗ is a correlated equilibrium

• Examples in the Chicken game:
• (Stay, Swerve) is a NE, and it can be a correlated equilibrium.
• Any of the public or private randomizations we saw before are correlated equilibria, but they
aren’t NE.
Equilibrium Refinements in Strategic-form games
(Technical)
• Mixed strategies can help us discard NEs which seem fragile to small
strategic mistakes, as if a player’s hands could “tremble” when choosing
her strategy. Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 1,1 0,0
Player 1
𝐷𝐷 0,0 0,0

• The above game has two psNE: (U,l) and (D,r).


• The second one seems more fragile to trembles:
• if player 1 deviates from D to U, even if U only occurs with a small probability, player
2’s BR would change from r to l.
• A similar argument applies if player 2 deviates from r to l, by a small prob.
• The issue, of course, is that in (D,r) players use weakly dominated
strategies.
• We next seek to rule out psNEs that aren’t robust to trembles.
Equilibrium Refinements in Strategic-form games
(Technical)
• Definition. Totally mixed strategy.
• Player 𝑖𝑖 ′ 𝑠𝑠 mixed strategy, 𝜎𝜎𝑖𝑖 , is “totally mixed” if it assigns a strictly
positive probability weight on every pure strategy, that is 𝜎𝜎𝑖𝑖 𝑠𝑠𝑖𝑖 > 0
for all 𝑠𝑠𝑖𝑖 .

• Therefore, all pure strategies happen, even with small probability.


• This allows for trembles, where D could occur with 0.001 probability
or less.
Equilibrium Refinements in Strategic-form games
(Technical)
• Definition. Trembling-Hand Perfect equilibrium.
• A mixed strategy profile 𝜎𝜎𝑖𝑖 = 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 is a Trembling-Hand Perfect Equilibrium
(THPE) if:
1. There∞exists a sequence of totally mixed strategies for each player 𝑖𝑖,
𝜎𝜎𝑖𝑖𝑘𝑘 𝑘𝑘=1 , that converges to 𝜎𝜎𝑖𝑖 , and
𝑘𝑘
2. for which 𝜎𝜎𝑖𝑖 ∈ 𝐵𝐵𝐵𝐵𝑖𝑖 𝜎𝜎−𝑖𝑖 for every 𝑘𝑘.

• Informally, these two requirements say that:


1. Every player i’s totally mixed strategy (which allows for trembles) must converge to 𝜎𝜎𝑖𝑖 ;
and
𝑘𝑘
2. Strategy 𝜎𝜎𝑖𝑖 is player i’s BR to her rivals’ strategy profile 𝜎𝜎−𝑖𝑖 at every point of the
sequence (i.e., for all k).
• Second requirement is a bit trickier to show. (Example in a moment.)
Properties of THPE
1. Every THPE must be a NE.
2. Every strategic-form game with finite strategies for each player has a
THPE.
3. Every THPE assigns zero probability weight on weakly dominated
strategies.
Intuitively, points (1) and (2) show that THPEs are a subset of the set of all
NEs in a strategic-form game.

𝜎𝜎 is a THPE ⇒ 𝜎𝜎 is a NE

And point (3) helps us rule out strategies D for player 1 and r for player 2 in
the 2x2 game we used as a motivation. Therefore, (D,r) is a NE but cannot be
supported as a THPE.
Example 5.9. Trembling-hand Perfect Equilibrium
Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 1,1 0,0
Player 1
𝐷𝐷 0,0 0,0
Matrix 5.14. A Game with two psNEs, but only 𝑈𝑈, 𝑙𝑙 is THPE

• Consider the following sequence of totally mixed strategies


𝜀𝜀𝑘𝑘 𝜀𝜀𝑘𝑘 1
𝜎𝜎𝑖𝑖𝑘𝑘 = 1 − , for every player 𝑖𝑖, where 𝜀𝜀𝑘𝑘 = .
2 2 2𝑘𝑘
• Example:
1 3 1 1
• When 𝑘𝑘 = 1, 𝜀𝜀1 = ,and 𝜎𝜎𝑖𝑖𝑘𝑘 becomes 𝜎𝜎𝑖𝑖1 = , , indicating that every player 𝑖𝑖 makes mistakes with
2 4 4 4
probability.
1 7 1
• When 𝑘𝑘 = 2, 𝜀𝜀2 = ,and 𝜎𝜎𝑖𝑖𝑘𝑘 becomes 𝜎𝜎𝑖𝑖2 = , , representing that mistakes are now less likely.
4 8 8
• In the limit, we find that (see figure in next slide)
1
lim 𝜎𝜎𝑖𝑖𝑘𝑘 = 1,0 since lim 𝜀𝜀𝑘𝑘 = lim =0
𝑘𝑘→+∞ 𝑘𝑘→+∞ 𝑘𝑘→+∞ 2𝑘𝑘
• which implies that player 1 (2) chooses U (l, respectively) in pure strategies, yielding strategy
profile 𝑈𝑈, 𝑙𝑙 .
Properties
• Generally, as 𝑘𝑘 increase, mistakes become less likely, and the above
totally mixed strategy converges to the psNE 𝑈𝑈, 𝑙𝑙 .
• This leads to the following figure:
Example 5.9. Trembling-hand Perfect
Equilibrium
• Therefore, the NE 𝑈𝑈, 𝑙𝑙 can be supported as a THPE because:
1. The totally mixed strategy 𝜎𝜎1𝑘𝑘 𝜎𝜎2𝑘𝑘 converges to 𝑈𝑈 (𝑙𝑙) ; and
2. 𝑈𝑈 (𝑙𝑙) is the best response of player 1 (2) to her rival’s totally mixed strategy,
𝜎𝜎2𝑘𝑘 𝜎𝜎1𝑘𝑘 , 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 for all k.
• To see point (2), note that:
3 1
• When k=1, 𝜎𝜎2𝑘𝑘 becomes 𝜎𝜎21 = , , where U is player 1’s best response because
3 1 3 4 4 3 1
𝐸𝐸𝐸𝐸1 𝑈𝑈|𝜎𝜎2 = 1 + 0 = and 𝐸𝐸𝐸𝐸1 𝐷𝐷|𝜎𝜎21 = 0 + 0 = 0.
1
4 4 4 4 4
𝑘𝑘 2 7 1
• When k=2, 𝜎𝜎2 becomes 𝜎𝜎2 = , , and U is still player 1’s best response because
7 1 7 8 8 7 1
𝐸𝐸𝐸𝐸1 𝑈𝑈|𝜎𝜎22 = 1 + 0 = and 𝐸𝐸𝐸𝐸1 𝐷𝐷|𝜎𝜎21 = 0 + 0 = 0.
8 8 8 8 8
𝑘𝑘 𝜀𝜀 𝜀𝜀 𝜀𝜀
• Same argument applies to every k since 𝐸𝐸𝐸𝐸1 𝑈𝑈|𝜎𝜎2 = 1 − 𝑘𝑘 1 + 𝑘𝑘 0 = 1 − 𝑘𝑘 > 0 and
𝜀𝜀𝑘𝑘 𝜀𝜀𝑘𝑘 2 2 2
𝑘𝑘
𝐸𝐸𝐸𝐸1 𝐷𝐷|𝜎𝜎2 = 1 − 0 + 0 = 0.
2 2
• Same argument applies to player 2’s best response to 𝜎𝜎1𝑘𝑘 being l for every k. (Check as a
practice.)
Example 5.9. Trembling-hand Perfect
Equilibrium
• In contrast, (D,r) cannot be sustained as THPE.
• While we can find converging sequences of totally mixed strategies (first
requirement)…
• Choosing D (r) is not player 1’s (2’s) best response to her rival’s totally mixed
strategy for every k (second requirement).
• To see this point, consider this totally mixed strategy:
𝜀𝜀𝑘𝑘 𝜀𝜀𝑘𝑘 1
𝜎𝜎𝑖𝑖𝑘𝑘 = ,1 − for every player 𝑖𝑖, where 𝜀𝜀𝑘𝑘 = .
2 2 2𝑘𝑘

• which assigns the opposite probability weights than that converging


to (U,l).
• It converges to psNE (D,r). Check!
Example 5.9. Trembling-hand Perfect
Equilibrium
• However, U is player 1’s BR to 𝜎𝜎2𝑘𝑘 for every k.
• To see this point, consider that:
1 3
• When k=1, 𝜎𝜎2𝑘𝑘 becomes 𝜎𝜎21 = , , and U is player 1’s best response.
4 4
1 7
• When k=2, 𝜎𝜎2𝑘𝑘 becomes 𝜎𝜎22 = , , and U is still player 1’s best response.
8 8
• Same argument applies for every k.
• Recall that finding that U is player 1’s BR, instead of D, for at least one value of
k and for at least one player would have been enough to show that (D,r)
cannot be sustained as THPE.
𝜀𝜀 −Proper Equilibrium
• THPE helps us rule out NEs that aren’t robust to trembles.
• But, which trembles do we allow?
• Myerson (1978) suggested that a rational player, while making
mistakes, should put:
• Higher probability weight on strategies yielding higher payoffs.
• Lower probability weight on strategies yielding lower payoffs.
• Alternatively, players are less likely to make costly mistakes.
𝜀𝜀 −Proper Equilibrium
• Definition. 𝜺𝜺 − proper equilibrium. For any 𝜀𝜀 > 0, a totally mixed strategy
profile 𝜎𝜎 = 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 is the 𝜀𝜀 − proper equilibrium if, for every player 𝑖𝑖, and
for every two pure strategies 𝑠𝑠𝑖𝑖 ,and 𝑠𝑠𝑖𝑖′ ≠ 𝑠𝑠𝑖𝑖 such that
𝑢𝑢𝑖𝑖 𝑠𝑠𝑖𝑖 , 𝜎𝜎−𝑖𝑖 > 𝑢𝑢𝑖𝑖 𝑠𝑠𝑖𝑖′ , 𝜎𝜎−𝑖𝑖 ,
• we must have that probabilities of playing 𝑠𝑠𝑖𝑖 and 𝑠𝑠𝑖𝑖′ , 𝜎𝜎𝑖𝑖 𝑠𝑠𝑖𝑖 and 𝜎𝜎𝑖𝑖 𝑠𝑠𝑖𝑖′ satisfy
𝜀𝜀 × 𝜎𝜎𝑖𝑖 𝑠𝑠𝑖𝑖 ≥ 𝜎𝜎𝑖𝑖 𝑠𝑠𝑖𝑖′

• Intuitively, if player 𝑖𝑖 ′ 𝑠𝑠 expected payoff from choosing 𝑠𝑠𝑖𝑖 is higher than that
from 𝑠𝑠𝑖𝑖′ , then…
• The probability of playing 𝑠𝑠𝑖𝑖 must be at least “𝜀𝜀 times higher” than the probability of
playing 𝑠𝑠𝑖𝑖′ .
Example 5.10. 𝜀𝜀 −Proper Equilibrium
Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 1,1 0,0
Player 1
𝐷𝐷 0,0 0,0
Matrix 5.14. A Game with two psNEs, but only 𝑈𝑈, 𝑙𝑙 is THPE

𝜀𝜀 𝜀𝜀
• Consider 𝜎𝜎𝑖𝑖 = 1 − , for every player 𝑖𝑖, where 𝑎𝑎 ≥ 2 and 0 < 𝜀𝜀 < 1.
𝑎𝑎 𝑎𝑎
• This mixed strategy is an 𝜀𝜀 − proper equilibrium because: (1) it is a totally
mixed strategy, assigning a positive probability weight to all players’
strategies; and (2) for pure strategies 𝑈𝑈 and 𝐷𝐷, their expected utilities
satisfy 𝜀𝜀 𝜀𝜀 𝜀𝜀
𝑢𝑢1 𝑈𝑈, 𝜎𝜎2 = 1 1 − + 0 = 1 − > 0 = 𝑢𝑢1 𝐷𝐷, 𝜎𝜎2
𝑎𝑎 𝑎𝑎 𝑎𝑎

𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 2 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 2
𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑙𝑙 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑟𝑟
Example 5.10. 𝜀𝜀 −Proper Equilibrium Example
And the probabilities of applying 𝑈𝑈 and 𝐷𝐷 are
𝜀𝜀 𝜀𝜀(𝑎𝑎−𝜀𝜀)
𝜀𝜀 × 𝜎𝜎1 𝑈𝑈 = 𝜀𝜀 1 − = and
𝑎𝑎 𝑎𝑎
𝜀𝜀
𝜎𝜎1 𝐷𝐷 =
𝑎𝑎
which satisfy
𝜀𝜀(𝑎𝑎−𝜀𝜀) 𝜀𝜀
𝜀𝜀 × 𝜎𝜎1 𝑈𝑈 = ≥ = 𝜎𝜎1 𝐷𝐷
𝑎𝑎 𝑎𝑎
since, after rearranging, this inequality simplifies to 𝑎𝑎 ≥ 𝜀𝜀, which
holds given that 𝑎𝑎 ≥ 2 and 0 < 𝜀𝜀 < 1 by assumption.
(Since the game is symmetric, a similar argument applies to player 2’s
utility from choosing 𝑙𝑙 and 𝑟𝑟, and its associated probabilities.
Proper Equilibrium
• Definition. Proper Equilibrium. A mixed strategy profile 𝜎𝜎 = 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 is a
proper equilibrium if there exists:

1. A sequence 𝜎𝜎𝑖𝑖𝑘𝑘 𝑘𝑘=1
that converges to 𝜎𝜎𝑖𝑖 for every player 𝑖𝑖

2. A sequence 𝜀𝜀𝑖𝑖𝑘𝑘 𝑘𝑘=1
where 𝜀𝜀 𝑘𝑘 > 0 for all 𝑘𝑘, that converges to zero

3. 𝜎𝜎𝑖𝑖𝑘𝑘 𝑘𝑘=1
is an 𝜀𝜀𝑘𝑘 −proper equilibrium for every 𝑘𝑘

• Proper equilibrium are also THPE, but the converse in not necessarily true.
• In other words:
• If 𝜎𝜎 is a proper equilibrium, it must be robust to a sequence of decreasing trembles
where costly mistakes are less likely to occur;
• while 𝜎𝜎 being THPE only requires that it is robust to any sequence of decreasing
trembles.
Example 5.11. Proper Equilibrium
Player 2
𝑙𝑙 𝑟𝑟
𝑈𝑈 1,1 0,0
Player 1
𝐷𝐷 0,0 0,0
Matrix 5.14. A Game with two psNEs, but only 𝑈𝑈, 𝑙𝑙 is THPE

• The sequence of totally mixed strategies from example 5.9


𝑘𝑘 𝜀𝜀𝑘𝑘 𝜀𝜀𝑘𝑘 1
𝜎𝜎𝑖𝑖 = 1 − , for every player 𝑖𝑖, where 𝜀𝜀𝑘𝑘 = 𝑘𝑘 ,
2 2 2
is a proper equilibrium if it satisfies the three requirements in the above
definition:
1. A sequence 𝜎𝜎𝑖𝑖𝑘𝑘 converges to 𝑈𝑈, 𝑙𝑙
2. 𝜀𝜀 𝑘𝑘 converges to zero
3. 𝜎𝜎𝑖𝑖𝑘𝑘 is an 𝜀𝜀𝑘𝑘 -proper equilibrium for every 𝑘𝑘 (as shown in Example 5.10).
Appendix
Fixed-point theorems, an Introduction
• Consider a function 𝑓𝑓: 𝑋𝑋 → 𝑋𝑋,
mapping elements from 𝑋𝑋 into X,
where 𝑋𝑋 ⊂ ℝ𝑁𝑁 .
• We then say that a point 𝑥𝑥 ∈ 𝑋𝑋 is a
“fixed point” if 𝑥𝑥 ∈ 𝑓𝑓(𝑥𝑥). For instance,
if 𝑋𝑋 ⊂ ℝ, we can define the distance
function
𝑔𝑔 𝑥𝑥 = 𝑓𝑓 𝑥𝑥 − 𝑥𝑥,
• which graphically measures the
distance from 𝑓𝑓 𝑥𝑥 to the 45-degree
line, as illustrated in Figure 5.8a and
5.8b.
Fixed-point theorems, an Introduction
• At points such as 𝑥𝑥 ′ where 𝑔𝑔(𝑥𝑥 ′ ) >
0, we have that 𝑓𝑓(𝑥𝑥 ′ ) > 𝑥𝑥 ′ ,
meaning that 𝑓𝑓(𝑥𝑥 ′ ) lies above the
45-degree line.
• In contrast, at points 𝑥𝑥 ′′ > 𝑥𝑥 ′
where 𝑔𝑔(𝑥𝑥 ′′ ) < 0, we have that
𝑓𝑓(𝑥𝑥 ′′ ) < 𝑥𝑥 ′′ , entailing that 𝑓𝑓(𝑥𝑥 ′′ )
lies below the 45-degree line.
Fixed-point theorems, an Introduction
• Since 𝑥𝑥 ′′ > 𝑥𝑥 ′ , if the distance function is continuous, we can invoke the
intermediate value theorem to say that:
• There must be an intermediate value, 𝑥𝑥, � between 𝑥𝑥 ′ and 𝑥𝑥 ′′ (or more than one)
where 𝑔𝑔 𝑥𝑥� = 0, implying that
• 𝑔𝑔 𝑥𝑥� = f 𝑥𝑥� − 𝑥𝑥� = 0, or f 𝑥𝑥� = 𝑥𝑥,
� as required for a fixed point to exist.
• Note that if 𝑓𝑓(𝑥𝑥) was not continuous, then:
• 𝑔𝑔(𝑥𝑥) would not be continuous either…
• Allowing for 𝑔𝑔 𝑥𝑥 ′ > 0 and 𝑔𝑔 𝑥𝑥 ′′ < 0 to occur,
• yet we could not guarantee the existence of an intermediate point 𝑥𝑥� between 𝑥𝑥 ′
and 𝑥𝑥 ′′ for which 𝑔𝑔 𝑥𝑥� = 0.
• Brouwer’s fixed-point theorem formalizes this result (next slide).
Brouwer’s fixed-point theorem
Definition. Brouwer’s fixed-point theorem.
If 𝑓𝑓: 𝑋𝑋 → 𝑋𝑋 is a continuous function, where 𝑋𝑋 ⊂ ℝ𝑁𝑁 , then it has at
least one fixed point, that is, a point 𝑥𝑥 ∈ 𝑋𝑋 where 𝑓𝑓 𝑥𝑥 = 𝑥𝑥.
• While Brouwer’s fixed-point theorem is useful when dealing with best
response functions, it does not apply to best response correspondences.
• where player 𝑖𝑖 is, for instance, indifferent between two or more of her pure strategies
when her opponent chooses strategy 𝑠𝑠𝑗𝑗 .
• The following theorem generalizes Brouwer’s fixed-point theorem to
correspondences.
• For a more detailed presentation on fixed-point theorems, see Border (1985).
Kakutani’s fixed-point theorem
Definition. Kakutani’s fixed-point theorem.
A correspondence 𝐹𝐹: 𝑋𝑋 → 𝑋𝑋, where 𝑋𝑋 ⊂ ℝ𝑁𝑁 , has a fixed point, that is,
a point 𝑥𝑥 ∈ 𝑋𝑋 where 𝐹𝐹 𝑥𝑥 = 𝑥𝑥, if these conditions hold:
1. 𝑋𝑋 is a compact, convex, and non-empty set.
2. 𝐹𝐹(𝑥𝑥) is non-empty.
3. 𝐹𝐹(𝑥𝑥) is convex.
4. 𝐹𝐹(𝑥𝑥) has a closed graph.
Nash Existence theorem
• First, define player 𝑖𝑖 ′ 𝑠𝑠 pure strategy set, 𝑆𝑆𝑖𝑖 , to be finite, i.e., a discrete
list of pure strategies, and denote a mixed strategy for this player as
𝜎𝜎𝑖𝑖 , where 𝜎𝜎𝑖𝑖 ∈ ∑𝑖𝑖, meaning that player 𝑖𝑖 chooses her randomization
among all possible mixed strategies available to her. Therefore, the
Cartesian product
∑𝑖𝑖× ∑−𝑖𝑖 = ∑
denotes the set of all possible mixed strategy profiles in the game, so
that every strategy profile 𝜎𝜎 = 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 satisfies 𝜎𝜎 ∈ ∑ .
• Second, let us define player 𝑖𝑖 ′ s best response correspondence to her
rivals choosing 𝜎𝜎−𝑖𝑖 as 𝜎𝜎𝑖𝑖 ∈ 𝐵𝐵𝐵𝐵𝑖𝑖 (𝜎𝜎−𝑖𝑖 ).
Nash Existence Theorem
• We now define the joint response correspondence 𝐵𝐵𝐵𝐵 𝜎𝜎 , as the product
of 𝐵𝐵𝐵𝐵𝑖𝑖 𝜎𝜎−𝑖𝑖 and 𝐵𝐵𝐵𝐵−𝑖𝑖 𝜎𝜎𝑖𝑖 , that is,
𝐵𝐵𝐵𝐵 𝜎𝜎 ≡ 𝐵𝐵𝐵𝐵𝑖𝑖 𝜎𝜎−𝑖𝑖 × 𝐵𝐵𝐵𝐵−𝑖𝑖 𝜎𝜎−𝑖𝑖
• Importantly, if 𝐵𝐵𝐵𝐵 has a fixed point, then, a NE exists.
• Therefore, we next check if 𝐵𝐵𝐵𝐵 satisfies the four conditions on Kakutani’s
fixed-point theorem, as that would guarantee the existence of a NE.

• Before doing that, we identify:


• 𝑋𝑋 in Kakutani’s fixed-point theorem with the set of all possible mixed strategy
profiles, ∑ , and
• correspondence 𝐹𝐹 with 𝐵𝐵𝐵𝐵.
Nash Existence Theorem
1. ∑ is a non-empty, compact, and convex set.
a. The set ∑ is non-empty as long as players have some strategies, so we can
identify pure or mixed strategy profiles.
b. Recall that if a set is closed and bounded, it is compact. The set of all
possible mixed strategy profiles is closed and bounded, thus satisfying
compactness.
c. Convexity is satisfied since:
• For any two strategy profiles, 𝜎𝜎 and 𝜎𝜎 ′ ,
• their linear combination 𝜆𝜆𝜆𝜆 + 1 − 𝜆𝜆 𝜎𝜎 ′ where 𝜆𝜆 ∈ 0,1 ,
• is also a mixed strategy profile, thus being part of ∑ .
Nash Existence Theorem
2. 𝐵𝐵𝐵𝐵 𝜎𝜎 is nonempty.
Since every player 𝑖𝑖 ′ 𝑠𝑠 payoff, 𝑢𝑢𝑖𝑖 (𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 ), is linear in both 𝜎𝜎𝑖𝑖 and 𝜎𝜎−𝑖𝑖 (expected utility is
linear in the probabilities)…
She must find a maximum (a best response to her rivals choosing 𝜎𝜎−𝑖𝑖 ) among her
available strategies, ∑𝑖𝑖 , which we know it is a compact set from point 1b.
Because 𝜎𝜎𝑖𝑖 ∈ 𝐵𝐵𝐵𝐵𝑖𝑖 (𝜎𝜎−𝑖𝑖 ) and 𝜎𝜎−𝑖𝑖 ∈ 𝐵𝐵𝐵𝐵𝑖𝑖 (𝜎𝜎𝑖𝑖 ) are non-empty (best response exists),
then their product, 𝐵𝐵𝐵𝐵 𝜎𝜎 , must also be non-empty.
3. 𝐵𝐵𝐵𝐵 𝜎𝜎 is convex.

To prove this point, consider two strategies for player ′
𝑖𝑖, 𝜎𝜎𝑖𝑖 and 𝜎𝜎 𝑖𝑖 , that are best
responses to her rivals choosing 𝜎𝜎−𝑖𝑖 , that is, 𝜎𝜎𝑖𝑖 , 𝜎𝜎𝑖𝑖 ∈ 𝐵𝐵𝐵𝐵𝑖𝑖 (𝜎𝜎−𝑖𝑖 ) .
Because both 𝜎𝜎𝑖𝑖 and 𝜎𝜎𝑖𝑖′ are best responses, they must both yield the same expected
payoff; otherwise, one of them cannot be a best response.
Therefore, a linear combination of 𝜎𝜎𝑖𝑖 and 𝜎𝜎𝑖𝑖′′ , 𝜆𝜆𝜎𝜎𝑖𝑖 + 1 − 𝜆𝜆 𝜎𝜎𝑖𝑖′ where 𝜆𝜆 ∈ 0,1 , must
yield the same expected

payoff as 𝜎𝜎𝑖𝑖 and 𝜎𝜎𝑖𝑖 , thus being a best response as well, that is,
𝜆𝜆𝜎𝜎𝑖𝑖 + 1 − 𝜆𝜆 𝜎𝜎𝑖𝑖 ∈ 𝐵𝐵𝐵𝐵𝑖𝑖 (𝜎𝜎−𝑖𝑖 ) .
Nash Existence Theorem
4. 𝐵𝐵𝐵𝐵 𝜎𝜎 has a closed graph.
This property means that the set 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 | 𝜎𝜎𝑖𝑖 𝜖𝜖 𝐵𝐵𝐵𝐵𝑖𝑖 𝜎𝜎−𝑖𝑖 is
“closed,” meaning that every player 𝑖𝑖 ′ 𝑠𝑠 best response correspondence
has no discontinuities.
The best responses depicted in this chapter, for instance, showed no
discontinuities. Because every player 𝑖𝑖 ′ 𝑠𝑠 payoff, 𝑢𝑢𝑖𝑖 (𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 ), is
continuous and compact, the set 𝜎𝜎𝑖𝑖 , 𝜎𝜎−𝑖𝑖 | 𝜎𝜎𝑖𝑖 𝜖𝜖 𝐵𝐵𝐵𝐵𝑖𝑖 𝜎𝜎−𝑖𝑖 is closed.
Nash Existence Theorem
• The previous 4 properties guarantee that a NE exists when players
face finite strategy spaces (i.e., a list of pure strategies).

• What if they choose their strategies from a continuous strategy space,


as when firm set their prices or output levels?
• Glicksberg (1952) extended the above result to setting with continuous
strategy spaces, where 𝑆𝑆𝑖𝑖 ⊂ ℝ𝑁𝑁 , showing that, if:
1. Every player 𝑖𝑖 ′ 𝑠𝑠 strategy space, 𝑆𝑆𝑖𝑖 , is compact, and
2. Her utility function, 𝑢𝑢𝑖𝑖 . , is continuous,
• Then, a NE exists, in pure or mixed strategies.
• (For generalization of this result to non-continuous utility functions, see
Dasgupta and Maskin (1986).)

You might also like